current position:Home page > assistant >

[Michael Thomas Madden 21]CVPR 2021 draw!The highest award has no Chinese figure,The first Huang Tao Memorial Award


Big data collection

Just just,As one of the three top clubs in computer visual fields,CVPR announced this year’s best papers, the best student papers and other award winners.

can be seen,Unlike Chinese, the best papers and best student papers last year are different.This year, the performance of the Chinese is a lot of “low-key”.In the best papers, the best student papers, etc.Only the best student papers have Chinese participation.

However, in the entire nomination list,In the 32 candidate papers of this CVPR 2021 Best Paper Award,There are 18 papers in Chinese to participate.Chinese one is more than 16 papers,Half of occupies.

In addition,The domestic agency scholars have a total of 6 papers.

According to the official report,This year received 7015 effective contributions,1663 papers are received,The received rate is 23.7%.

This data has risen slightly compared with 22% of the receipt last year.But you can think about it,Competition is still quite fierce.

Next,Let’s take a look at the specific awards together with the article.

Best paper

This year’s best papers is Giraffe: Representing Scenes As Compositional Generative Neural Feature Fields,Author is an autonomous vision group for the Map Smart System (Autonomous Vision Group).AVG) Ph.D. Michael Niemeyer in computer vision / machine learningAnd AVG leader, Professor of Computer Science Andreas Geiger.

In the paper,Researchers put forward,Depth generation model Although the photo-level realistic image can be synthesized at high resolution,But for many applications, it is not enough.Content creation also needs to be controllable.

Although there are research in recent years to discuss how to unlock potential changes in data,But they only consider two-dimensional space.In addition,Only a few studies have taken into account the combination of scenes.

therefore,The researchers expressed the scene as the synthetic neurotransmitting field.Make it to separate one or more targets and the shape and appearance of a single target from the background.At the same time, you will learn from unstructured and unused image collections without any additional supervision.

This is also the key assumption that this article is,That is, the synthesized three-dimensional scenarios will be merged into the generation model to result in more controllable image synthesis.

This scene represents the combination of the nerve rendering Pipeline,A fast and realistic image synthesis can be obtained.Experiment proves thatThis model can separate a single target object,And allow you to pan and rotate them and change the camera posture in the scene.

Paper address:

Https: // arxiv.ORG / PDF / 2011.12100.PDF

Best Paper Honor Prize:

Best student papers

This year’s best student papers are Task Programming: Learning Data Efficient Behavior RePresentation,The author is Jennifer J, California University of Technology. Ann Kennedy et al. Ann Kennedy, Assistant Professor of Northwestern University.

In the paper,Researchers proposed a valid trajectory embedding method for behavioral analysis Treba,Experts can effectively design tasks through “task programming”.At the same time, you can construct a small amount of programming tasks by switching data annotation times.reduce workload.

we know,On the labeling data set,To achieve accuracy,In addition to the necessary expertise,Need to pay a lot of time and energy,This increases the work burden for experts in invisibular.Especially in the automatic behavior analysis, it is very obvious in Automated Behavior Analysis.For example, depel from video tracking data or the like.

therefore,In order to reduce the workload of comments,The researchers put forward this effective trajectory embedding method based on multitasking self-supervision learning TREBA.

The researchers used a data set in the field of behavioral neuroscience to evaluate the method.Test of three data sets in both mice and fruit flies,the result shows,By using TREBA embedding,The annotation burden is reduced by 10 times.That is to say,Task planning and self-supervision is an effective way to reduce the amount of detailed information on the field expert.

Paper address:

Best Student Paper Honor Prize:

Pamitc Award

The two winners of the Pami Youth Researcher Award this year are Georgia Gkioxari and Phillip Isola.

Georgia Gkioxari is a scientist of Facebook Fair,Doctoral graduated from the University of California Berkeley.The main research field of Gkioxari is computer vision.It is also one of the developers of Pytorch3D, one of the author of Mask R-CNN,There are currently 16,000 cited.


Phillip isola is Assistant Professor Mit EECS,Mainly study computer vision, machine learning and artificial intelligence.Ph.D. graduated from MIT’s brain and cognitive science major,It is also a postdoctoral after the University of California, Berkeley.I have a year of access to research scientists in Openai.

At present, Isola’s paper reference number is 28056,Which is the highest paper as “Image-to-Image Translation with Conditional Adversarial Networks”,This paper studies the image of the condition to image conversion,It can be said that it is a front stage of CYCLEGAN.


This year’s classic papers (Longuet-Higgins Award) comes from Microsoft and Shixi University.

Microsoft’s paper entitled “Real-Time Human Pose Recognition in Parts from Single Depth Images”,Published in CVPR 2011,Quote 4108.

This paper puts forward a kind of in the absence of time information without using time information.A method of quickly accurately predicting the 3D position of the body joint from a single depth image.

Paper link:

The paper of Shixi University is the Baby Talk: Understanding and Generating Simple Image Descriptions,Published in CVPR 2011,Current quote 1159.

In the paper,The researchers are based on a lot of language data.A system that automatically generates a natural language from an image is proposed.The system utilizes statistics collected from parsing a large amount of text data and computer visual identification algorithms.Very effective in generating sentences for images,It is an important job of early images to text.

Paper link:

Thomas S. Huang Memorial Award

A award has also been added this year.This is Thomas S. Huang Memorial Award.

last year,In order to remember a generation of CV masters, 84-year-old Chinese computer visual Thai THOMAS S. Huang (Huang Tao),The CVPR conference established THOMAS S. Huang Memorial Award.The winners of the award will be elected by the Pamitc Committee.

The first Thomas S. The winner of the Huang Memorial Award is MIT Electronic Engineering and Computer Science Professor Antonio Torralba.

Torralba’s research sector includes scene understanding and context-driven target identification, multi-sensory perceptual integration, data set construction, and visualization and explanation of neural network characterization.

His current paper quote is 78736,The H index is 111.


Https: // groups.CSAIL.Mit.EDU / Vision / Torralbalab /

Chinese figure in this CVPR

Although in the final prize,It doesn’t see the papers of Chinese authors.But still undeniable,Chinese people have more and more appearance on the AI top.

Take CVPR 2021 as an example,In the candidate list of the best papers,32 candidate papers,18 articles have Chinese participation,The authors come from universities and research institutions at home and abroad.

There are 16 papers in Chinese.Among them, there is a famous scholar of Potae, Shen Chunhua, and He Wei Ming.

In addition to the colors of the paper,In this year’s CVP review,Professor Tan Tie Culch, Chinese Academy of Sciences, served as Chairman of the General Assembly,Yan Jingyi, Professor, Shanghai University of Science and Technology, the Lifelong Professor Yang Ruigang, a Lifelong Professor of the Kentucky University, served as Chairman.

I believe in the future,Chinese body will become an important impact of an AI academic world.